Dylan Di Filippo
Dylan Di FilippoAI journal

Journal IA 2026-W39

Understand what changes. Choose what deserves a test.

Understand what changes.

Public sources. Verified facts. Analysis kept separate from announcements.

5 stories

GoogleVoice & creativity

Voice generation moves from choosing a voice to directing it

Google adds customizable voices and fine performance control. Narration, games and dubbing become more directly manageable uses.

Deployment API / AI Studio

Official Gemini Voice Generation Visual
Google: official visual from the pilot
Facts, analysis and limits

Verified facts

  • Gemini 3.8 Flash TTS lets you design voices with natural-language instructions and direct pacing, emotion and acting line by line.
  • Flash-Lite TTS targets high-volume and cost-sensitive uses. AI Studio and the Gemini API are being deployed; the Gemini Enterprise API is announced later.

What changes

For creative content, accent and emotion become setting parameters rather than fixed presets.

Practical implication

To be tested on a short original narrative, without cloning of real person.

Availability

Belgium: voice replication in AI Studio unavailable (EEA restriction), to be distinguished from the creation of original voices. Gemini Enterprise: forthcoming. Prices to be checked according to the applicable grid.

Limits and uncertainty

The creation of a voice does not permit the imitation of a person without consent. The advertised quality has not been assessed here.

Read full article →
GoogleVoice & creativity

An assistant who speaks, appears and acts at the same time

Live Avatar adds a visual presence to voice agents. Immediate use is mainly aimed at companies, support and interactive guided tours.

Gemini Enterprise

Gemini Live Avatar: official Google demo. Official demonstration, not an independent test. Press play to start.
Facts, analysis and limits

Verified facts

  • Gemini 3.8 Live with Live Avatar combines audio conversation and video presence with lip and expression synchronization.
  • The system can call tools in the background during the conversation. Custom avatars require company permission.

What changes

The voice becomes an interface to perform tasks, not just read a response aloud.

Practical implication

Assess whether an avatar actually adds value over a simple voice interface.

Availability

Gemini Enterprise; personalized access subject to allowlisting. Not a new universal feature of the Gemini General Public app.

Limits and uncertainty

A smooth demo does not measure the success rate of tasks or the cost of an agent in production.

Read full article →
AnthropicResearch & society

A biological discovery, but not yet a medical application

Claude helps identify an unusual system; researchers test it experimentally. The value lies in combining large-scale exploration with human validation.

Preliminary result

Official scientific illustration of the ART enzyme system
Anthropic: official visual of pilot
Facts, analysis and limits

Verified facts

  • Anthropic describes an ART enzyme system associated with DNA repeats, identified by Claude agents in genomic data.
  • The campaign described mobilizes about 950 agents for 21 hours and 210 million tokens; human scientists carry out laboratory experiments.
  • The main function of the system remains unknown; a preprint is published. It is not a validated medical tool nor a new CRISPR usable.

What changes

A concrete example where AI helps to select scientific hypotheses; experimentation remains indispensable.

Practical implication

Read to understand the mechanism, without turning a preliminary result into a promise of treatment.

Availability

Research publication and preprint; no therapeutic offer or product ready to use.

Limits and uncertainty

Similarity to some CRISPR motifs is a scientific lead, not demonstrated functional equivalence.

Read full article →
GoogleResearch & society

AI in orbit: a hardware test, not yet a space cloud

Suncatcher explores a long-term direction for AI infrastructure. The news is an upcoming test and its technical challenges, not Google's data centers moving into space.

Prototype & research

Facts, analysis and limits

Verified facts

  • Google is preparing a first orbital test of Project Suncatcher to measure the operation of its TPUs in space conditions.
  • The project had been announced the previous year; the article of this period details the tests and obstacles of radiation, vibration, cooling and network.
  • Laser-connected computation constellations are a research perspective, not a commercial data centre already in service.

What changes

Infrastructure and energy become central subjects of the race at the AI; this announcement illustrates the long-term bets.

Practical implication

To be monitored; no changes to your projects.

Availability

Experimental program; not a cloud service available for your applications.

Limits and uncertainty

Solar energy does not remove the cost of launchings, thermal stresses, or maintenance.

Read full article →
MetaApplications

Meta wants to get the agent into what you see.

The glasses must give Muse the visual context of everyday life. Meta combines personal assistant, voice and portable equipment.

Future Announcements & Arrivals

Muse: official Meta demonstration. Official demonstration, not an independent test. Press play to start.
Facts, analysis and limits

Verified facts

  • At Connect, Meta announces Muse on its AI glasses in the coming months, with a voice mode that enables background work.
  • Meta announces new connectors and an e-mail address for agent; Muse Charm is presented with more details promised later in the year.

What changes

The bet is to reduce the need to open an application or describe a scene before asking for help.

Practical implication

To be monitored; not to buy equipment on the promise of a future function alone.

Availability

The arrival of Muse on the glasses is future. Belgian availability and price of all these functions not established in this pilot.

Limits and uncertainty

Muse had already been launched earlier in September; the novelties dealt with here concern Connect, not the original launch of Muse.

23 Sept. · recap on 24

Read full article →

Choose what deserves a test.

Public sources. Verified facts. Analysis kept separate from announcements.

1 stories

OpenAICoding

Vision Sol/Luna: update comparisons

This fix may explain recent screenshot errors. Avoid comparing an affected older version with a corrected version as though only the model had changed.

Published Corrigendum

Facts, analysis and limits

Verified facts

  • OpenAI fixes an image encoding bug that degrades the visual understanding of GPT-6 Sol and Luna in API and Codex, including the computer use.
  • The provider recommends restarting the affected evaluations and workflows using images.

What changes

Replay a small stable set of UI captures and browser tasks. Keep date, model and harness with results.

Practical implication

Re-launch visual tests that failed before September 25.

Availability

API/Codex fix, not a new model release.

Limits and uncertainty

A before/after result may reflect this patch rather than the only generation change in the model.

Read full article →